Hi everyone!

I really want the latest JAV movies to be subtitled right away. I use Subtitle-Edit and Whisper to create subtitles from Japanese to English and then translate them to Indonesian. My problem is that my native language is not English and when I use the Whisper-medium model, the results are not accurate. Does anyone know how to make Whisper more accurate?

Thanks

Use the last model: the turbo model or the large model. But Japanese is not a language very accurate and hard to translate. Good luck.

The AI translators have a really tough time with Japanese. I tried using large Whisper with a short, simple, old Eriko Miura movie and the result didn't make much sense. I tried another translator and Google and finally figured out what was happening. If I ever assemble a subtitle from the three translations, I'll post it here.

There is another subtitle site that does not allow AI translated subtitles to be posted. They call them garbage. They only allow human translated subtitles.

@nyangleh: Maybe you talk about OpenSubtitles. I've read a topic about this rule lately. Before AVSubtitles, and before AI, I was an amateur subtitles creator and I had hard time with some of their admins, that's why I quit.
From my point of view, ban AI with no distinction is a nonsense, but OpenSubtitles is big and old enough to afford it.

About your Japanese tests, keep in mind that automatic translators can't handle asian languages so easily. So you can't expect a perfect result, but if you "help" it a little, you can improve it.

One thing I always do first, is to choose a scene with classic dialogs. Choose a scene from its beginning (from the moment the scene starts, when the character knocks the door, etc...). So the translator could get the context of the scene.

Then, remove all the index and timestamps of this scene (this part is important).

95
00:06:21,860 --> 00:06:24,320
なんかいい酒のつまみねえかな。

96
00:06:28,080 --> 00:06:31,960
こんなもんじゃ酒のつまみにもなんねえや。

97
00:06:37,500 --> 00:06:39,900
ここにいたじゃねえかよ、つまみが。

becomes:

なんかいい酒のつまみねえかな。
こんなもんじゃ酒のつまみにもなんねえや。
ここにいたじゃねえかよ、つまみが。


Then try to translate it with Deepl or Google.
If the result makes sense, so your Japanese source is fine.
Else, if you see words like "car's carburetor" while the scene obviously takes place in bathroom, give up: your Japanese source quality is too bad.

Reading most descriptions of the Japanese movies, you will readily notice discrepancies in pronouns/genders. (No, I am not trying to be politically correct)

It would seem that in the Japanese languages there is no gender difference in the words used regarding people. It all depends on whom you are speaking to or about. The word for mother/father, same. He/she, same. Son/daughter, same. Haven't checked on aunt/uncle or brother/sister, but I could take a guess. AI translators simply can't handle/delineate this aberration, so a line by line will always be needed.

My only advice to non-English users would be to select those English subtitles that have earned perfect ratings and translate those into your native tongue, but considering how strange the English language is (Three genders, he, she, it), you'll probably need to do a line-by-line anyways.

There is no easy fix for this particular problem.


RK

Then, remove all the index and timestamps of this scene (this part is important).
Then try to translate it with Deepl or Google.

Hi Truc

I use Google translate extensively and I find it is not necessary to remove the index and time stamps. GT retains them faithfully while translating the text.

Hello!
The issue is not really about the timestamps themselves. Well, some years ago, I saw wird things like

25
00:02:20,000 --> 00:02:23,000
Blabla in french

becoming

twenty five
00:02:20,000 --> 00:02:23,000
Blablah in english

But this kind of problems seems to have disappear.

That being said, I still remove the timestamps because I think it helps Google or Deepl, or any AI to get more context.
Of course it depends on the dialogs, but I've noticed an enhanced result for monologues or things like that.

Recently, I try translations with AI and the result is pretty good. But I still struggle to master this kind of tool and censorship is here too...(!) ChatGPT won't translate sexual related words :( Grok is not so prude, and works pretty well (but from my point of view, deepl is still the best)

Grok is not so prude, and works pretty well (but from my point of view, deepl is still the best)

Do you use the free or the paid versions of these? I find the free Deepl has a greater size limitation than GT, which means translating in smaller chunks, so more time consuming.

I've never tried Grok, will do so and compare for the next one.

I use the free version, and you're right: Deepl's allowed chunks are more and more small. It becomes ridiculous. But since I remove the timestamps that's quite fine.
Some years ago, I made a script to count characters in order to fit ^^; ... I'd be happy to share it, but it's really Linux specific.

Two of my most crazy ideas for this site are:

1) add a translation engine => I can create something to get the most of translators but... I need a translator. Deepl monthly cost for a website like avsubtitles is 5€ + 20€ per million of characters
Without timestamps, average of a JAV subtitles file is 50.000 characters. So the maths are easy: one translation around 1€.
I really don't know how subtitle-cat can manage it. I could get more server and host my own translation engine, but translation engines one can install himself are too bad (bad results).
(note: subtitle cat doesn't use Deepl. I don't know which engine they use, but for sure it's not deepl).

2) add a transcription engine. I think about adding Whisper here, but I need more CPUs (GPUs to be correct) so it means more money (again!) but also some stupid questions like "which location" because of the european censorhsip :/

I am no internet professional, so I'm still stuck with these stupid cost/location questions.